Still WIP, figuring out why the ANLS performance is lower with shorter training. Might come from the DistributedSampler and local_rank in place of global_rank?
else, adds
docVQA finetuning task, from a checkpoint pretrained on IDL or not
docVQA eval task on the val set (the test set is only through their website)
Still WIP, figuring out why the ANLS performance is lower with shorter training. Might come from the DistributedSampler and local_rank in place of global_rank? else, adds